It was critical that lab as well as truck-shipment data be integrated and available to the analytics package. Feed tank properties needed to be calculated automatically with each new addition of raw materials from truck shipments.
To apply advanced control to a process, its basic equipment must be functioning properly -- otherwise, you risk wasting time and effort in the long run. So, staff surveyed all process instrumentation and loops to ensure that all instruments were problem-free, and loops were tuned to achieve optimum performance.
Before going online with the trial, we ran training sessions so operations personnel were comfortable enough with the new application to use it in the field.
The field trial was run on two batch processes -- each made different products. The output of Process 1 was an input into Process 2 (Figure 1). We used 18 input variables, 38 process variables and 4 output variables for the modeling and online analytics. (Many more process variables could have been included but we deliberately kept the scope small for the trial.) Data were collected at 1-min. intervals. We relied upon historical batches for analysis and model development across these two processes.
We created a separate model for each process stage -- defined uniquely by product, equipment and operation performed. This allowed inputs and outputs used in analysis to differ for each stage.
The tools used for offline development of the models required selection of historical batches as well as appropriate variables from historical data. The tools enabled comparing the results of the model against historical data to determine the model's accuracy or if a particular batch was an outlier and, so, shouldn't be used in developing the model.
Once suitable models were in place, online analytics tools were put to use via a web-based interface. Such an interface was important because some process specialists who would be working with the analytics were located throughout the world.
Figure 2 shows the main screen monitored by operators. It displays active batches for each process, along with an indication of any process faults detected. To investigate a detected fault, the operator simply selects the batch and is taken to another display (Figure 3), which provides the statistical charts for the selected batch. Whenever statistical values for the batch exceeded the upper limit (a value of one), the trend for the indicators appears outside of the green zone. The operator can select anywhere on the line to see on the left side of the display the list of contributing variables for this point in time, in order of greatest contribution. To further analyze the situation, the operator can select any one variable to get its individual trend for the selected batch, which is overlaid against the model developed for the particular process, along with the upper and lower limits for the variable (Figure 4). In the example shown, it is clear the variable is trending well above the output from the model and the acceptable variation, so the operator can address the situation as needed.
Finally, the operator can view on a separate trend the predicted end-of-batch quality with confidence limits (Figure 5).
The Rouen facility has realized numerous and ongoing benefits that started immediately. Some include:
• During a "train the trainer" class, the online analytics uncovered a fault in the actual process -- a previously undetected problem with the mass flow meter for a key component charged into the batch. This fault was going unnoticed with "traditional" monitoring systems. This revelation helped highlight the benefits of the technology and certainly got the attention of operations management!